OpenAI Says It Solved a 90-Year-Old Millennium Math Problem
The OpenAI Navier-Stokes proof claims to resolve one of mathematics’ most notoriously difficult open questions, a Millennium Prize Problem that had remained unsolved for roughly 90 years, according to a new announcement from OpenAI.
The result, produced by an internal OpenAI system, demonstrates that the dynamics of the Navier-Stokes equations governing fluid motion can develop what’s known as a singularity within finite time. OpenAI has shared both a detailed written proof and a formal verification of the result using Lean, a programming language specifically designed for mathematical proof verification.
Why the Navier-Stokes Problem Matters
The Millennium Prize Problems represent some of the deepest unresolved questions at the frontier of mathematics. Whether smooth three-dimensional fluid motion could theoretically break down had remained an open question for approximately nine decades.
OpenAI stated that a major goal behind this work is empowering scientists to advance research that benefits humanity broadly. To tackle the Navier-Stokes problem specifically, the company used an internal model described as significantly more capable than its previous GPT-6 Astra system, with OpenAI emphasizing the importance of informing the public about the current pace of AI progress.
What the Navier-Stokes Equations Actually Describe
The Navier-Stokes equations apply Newton’s second law of motion to describe how fluids move, treating fluid as a continuous medium rather than tracking individual molecules. These equations underpin practical applications including aircraft design, weather forecasting, and the study of blood flow through the human body.
A fundamental open question has long been whether this continuous approximation of fluid behavior could theoretically break down, specifically whether Navier-Stokes equations for three-dimensional incompressible fluid could develop a singularity, meaning fluid speeds growing without bound within a finite amount of time, even when starting from smooth initial motion. Because real fluids cannot physically move at infinite speed, such a breakdown would represent a genuine limitation in how these equations model reality.
The equations themselves date back to nineteenth-century work by Claude-Louis Navier and George Gabriel Stokes. In 1934, mathematician Jean Leray proved that solutions exist in a generalized sense, though whether those solutions always remained smooth became a central unresolved question. In 2000, the Clay Mathematics Institute formally designated the Navier-Stokes existence and smoothness problem as one of seven Millennium Prize Problems.
What the OpenAI Navier-Stokes Proof Actually Shows
OpenAI’s system produced both an analytical proof and a Lean formalization demonstrating that an initially smooth fluid at rest can develop a singularity within finite time. Throughout this process, the fluid experiences a smooth applied force while its total energy remains finite from initial rest through the eventual formation of the singularity, satisfying the official Millennium Prize problem statement.
The core solution involves a vortex, a spinning swirl of fluid that spirals inward while becoming increasingly elongated, somewhat resembling spaghetti. This central region shrinks while speeding up in a mathematically precise way that keeps its total energy finite, consistent with fundamental physical laws. The key technical challenge involved getting the equations to develop this breakdown naturally through the fluid’s own motion, rather than artificially introducing infinite force.
OpenAI Navier-Stokes Proof and the Millennium Problem
Since August 28, OpenAI has been training a new internal model exhibiting what the company describes as unprecedented benchmark performance, particularly in mathematics, with training and performance improvements still ongoing.
On September 1, after hearing rumors that two separate Millennium Prize problems had potentially been resolved elsewhere, and inspired by their own internal model’s recent performance leap, OpenAI launched an effort to evaluate the system against all remaining open Millennium Prize problems alongside several other high-impact mathematical challenges.
The company used a coordinated system of AI agents powered by its internal model, giving agents access to tools including cached internet access and code execution capabilities. Agents were organized into varying group sizes, with the specific group that ultimately produced the Navier-Stokes resolution involving roughly 10,000 concurrent agents working together, all operating under the same strict safeguards OpenAI applies to its frontier model evaluations, including active monitoring and isolation protocols.
An Unexpected Breakthrough Along the Way
Beyond the core Millennium Prize problems, OpenAI’s multiagent system was also assigned several related “easier” problems, including a similar blowup question involving the Euler equations, essentially the Navier-Stokes problem with viscosity removed entirely. The agents unexpectedly resolved this related question as well, specifically the unforced version with no external force applied. Roughly 100 agents worked together for approximately 50 hours to produce that Euler regularity disproof.
After witnessing that unexpected Euler solution, OpenAI’s team determined Navier-Stokes represented the most promising remaining problem to pursue, shifting resources away from other Millennium Problems to focus resources there instead, while prompting agents directly with the Euler resolution as a foundation.
Different agent groups were encouraged to explore diverse mathematical approaches independently. Periodically, OpenAI cross-pollinated insights between agent groups using Codex to consolidate the most valuable findings from each group’s intermediate results, an approach that ultimately guided the specific group that found the Navier-Stokes solution.
The Scale of the Computational Effort
The agents arrived at their final resolution on Saturday, September 5, roughly 88 hours after the first agents were launched. Formal Lean verification then took an additional 17 hours to complete using the GPT-6 Astra model.
Across all attempted problems combined, OpenAI’s agents sent a combined 4.9 million messages and consumed approximately 300 billion output tokens throughout the process. Specifically for resolving the OpenAI Navier-Stokes proof, agents sent 2.7 million messages and used roughly 130 billion output tokens, underscoring the massive computational scale involved in tackling one of mathematics’ longest-standing unsolved problems.